Broadcom has reportedly arranged more than $50 billion for OpenAI in connection with custom ASIC deals, as custom chip supply-chain roles shift across the industry.
The custom chip (ASIC) sector has seen a string of financing arrangements worth hundreds of billions of dollars in less than a month, and the way chip suppliers and their customers work together is being redefined.
According to a report by 36Kr, Broadcom is arranging more than $50 billion in financing for OpenAI's custom AI chips, though negotiations are still at an early stage. Less than a week earlier, a financing arrangement of up to $42 billion between Anthropic and Broadcom had been disclosed to pay for TPU compute leasing. Two leading model companies, one advancing the deployment of its own chips and the other expanding its rental of custom compute, have both tied multi-billion-dollar financing to ASICs.
The design side is equally active. Synopsys signed a multi-year IP agreement worth more than $1 billion with Amazon, and SEMIFIVE won a $52 million order to develop an AI inference ASIC. On October 6, Marvell raised its fiscal 2029 custom chip business revenue target to more than $12 billion. Customer investment in custom compute is extending beyond the main accelerator to design services, IP and companion chips.
## Financing arrangements pull suppliers into customers' capital structures
The Anthropic-Broadcom arrangement corresponds to compute spending that has already been committed. Documents disclosed on October 1 show that Anthropic's five-year TPU compute leasing commitment is about $125.2 billion, and the financing line of up to $42 billion could cover roughly one-third of it. Broadcom or its designated financing partners will provide the funds in the form of convertible debt, and Anthropic is not expected to issue the relevant notes before completing an IPO. The credit line is not cash Broadcom has already spent, but it pushes the relationship between the two companies beyond chip supply and procurement.
This batch of compute uses TPUs developed jointly by Google and Broadcom. In April, Anthropic announced that it would begin obtaining several gigawatts of additional TPU compute starting in 2027, and the current financing arrangement provides a funding path for that deployment. For Google, TPUs are extending from serving internal businesses to meeting large-scale demand from external model companies, which expands its shipment headroom. Anthropic, meanwhile, can increase its use of ASIC compute without first completing its own chip.
Independent accelerator companies have also found new buyers. On September 29, Cerebras and General Compute announced a multi-year agreement under which General Compute will purchase, deploy and operate Cerebras systems, then sell dedicated inference compute to customers. Services are expected to begin in the first quarter of 2027. The first application area is coding agents. Such programs must call models continuously to complete planning, writing, testing and revision, so the wait for each inference accumulates across the entire task.
General Compute had previously obtained a $400 million debt financing line from Upper90. That money is not the amount of this chip contract, but it shows that new types of cloud service providers are taking on hardware financing and operations. This lets downstream inference service providers buy compute without holding the entire set of equipment themselves, giving independent chipmakers another route to market without requiring every application company to become a direct hardware buyer.
A day earlier, Cerebras and Gimlet Labs announced a plan to deploy 100 megawatts of inference compute, with the first partner data center expected to come online later this year. Gimlet operates an inference platform that includes multiple types of chips, and the two companies had previously worked together to serve customers' private inference tasks. These two consecutive partnerships bring dedicated accelerators into cloud service channels. They also show that market expansion does not depend solely on in-house chip systems such as those of Google and Amazon.
## Shipment forecasts versus revenue reality
An industry research report released in late September projects that in the 2027 high-end cloud AI accelerator market, ASIC unit shipments will surpass GPUs. Specifically, it forecasts about 25.95 million high-end cloud AI accelerators shipped in 2027, of which about 15.26 million will be ASICs. That would be up about 111% year on year and account for nearly 60% of the total. However, this metric measures chip units and cannot be used to infer that ASIC revenue or compute power will exceed GPUs. The same forecast projects that Nvidia-related shipments will still grow about 18%. When incremental demand is large enough, both architectures can scale up at the same time.
| Segment | Recent major deal or partnership | Amount or scale | Timing |
|---|---|---|---|
| Financing arrangement | Broadcom arranges financing for OpenAI custom chips | Over $50 billion | Disclosed in early October |
| Financing arrangement | Broadcom-Anthropic convertible debt financing | Up to $42 billion | Disclosed October 1 |
| IP licensing | Synopsys-Amazon multi-year IP agreement | Over $1 billion | September 30 |
| Design services | SEMIFIVE takes on inference ASIC development for a US chip company | About $52 million | September 29 |
| Guidance | Marvell raises fiscal 2029 custom chip revenue target | Over $12 billion | October 6 |
*Note: Most of the above amounts are contract ceilings or target values and do not equal confirmed revenue or cash already paid.*
Large-scale TPU deployment is still driven by a small number of customers, while new partnerships for independent accelerators must pass through data center commissioning and customer migration. Financing and long-term agreements bring future demand into the chip supply chain earlier, but whether sustained procurement materializes still depends on whether downstream services can absorb this compute.
## In-house chips move from blueprints to racks
OpenAI is pushing its in-house chip effort into the deployment stage. In late September, hardware chief Richard Ho said in an interview that Jalapeño will first meet OpenAI's own compute needs and that external sales are not currently a priority. System details disclosed on October 2 show that the in-house accelerator uses AMD's EPYC Turin as its host CPU.
Under the disclosed division of labor, Broadcom handles chip implementation and networking-related work, while Celestica participates in board and rack system integration. The model company designs accelerators around its own business while continuing to use mature CPUs and specialist system supply chains. Expanding its in-house scope does not require redeveloping every part of the server.
When evaluating Jalapeño, OpenAI compared inference energy efficiency under the same response-speed requirements and made reducing data movement a key design direction. Control over its models, software and actual usage requirements lets the company adjust the architecture around changes in computation and memory access during inference. Ho also stressed that Jalapeño is programmable and can run different models. Previous tests included open-weight models such as DeepSeek and Kimi. Jalapeño does not permanently fix a single model in silicon, but supporting new model families still requires developing compute kernels and continuous optimization.
On September 25, Amazon announced a specific adaptation case. Real-time interactive AI platform Reactor worked with the AWS Neuron team to run the video generation model Rolling Forcing on a single Trainium2, achieving real-time generation at more than 16 frames per second. Unlike generating a complete video at once, such applications must output frames continuously and respond to user input, so latency requirements apply through the entire generation process. The team developed dedicated compute kernels to handle changes in data dimensions, memory access and cache management, and rearranged how the model executes across the chip's cores. The result did not provide large-scale commercial deployment volumes, and it cannot be directly converted into a cost advantage over GPUs. It does demonstrate the work required for ASICs to enter new applications: chipmakers must work with model and platform developers to resolve bottlenecks in real operation before paper performance can become a service users can actually use.
The Gimlet-Cerebras partnership, by contrast, combines different chips. Gimlet's platform can allocate tasks by inference stage, letting GPUs and Cerebras systems each handle the work they are best suited for, and then deliver services through a unified interface. For application developers, dedicated accelerators can be added beneath existing services without migrating the entire service out of the current GPU system at once. Dedicated chips may first win a segment of the inference workflow rather than the entire cluster.
Under OpenAI's previously announced plans, Jalapeño is scheduled to begin deployment by the end of the year, and the company has made clear it will continue to use accelerators from